bhyve Enabled the nVidia GPU passthrough inside a Linux virtual machine via QEMU accelerated with BHYVE.

[eseipi]

You are making several unreasonable oppositions here :
[...]
your seem to be one ideological, hard, non-collaborative position that does not produce technological progress.
[...]
It's not good. Wars are waged to defend ideologies, that is, uncompromising positions.

Sure, you can call me a Luddite as much as you want. But again, how do you plan to bypass the FreeBSD contribution guidelines? Or these guys, which would be even harder?

Code provenance — QEMU documentation

www.qemu.org

TL;DR:

Current QEMU project policy is to DECLINE any contributions which are believed to include or derive from AI generated content. This includes ChatGPT, Claude, Copilot, Llama and similar tools.

This policy does not apply to other uses of AI, such as researching APIs or algorithms, static analysis, or debugging, provided their output is not included in contributions.

 

[NapoleonWils0n]

The obstacle that has not been mentioned
is the legal team for any project you want to submit code to.

AI is notorious for being trained by infringing copyright material and other peoples code.

I would imagine that the Freebsd foundation has a legal department
and if they are any good they would advise not accepting any AI code due to the risk of lawsuit.

For example suppose AI code was accepted into the kernel
and it was then found that Claude had been trained on copyright code from vmware.

You would imagine that in the small print for Claude TOS they would have a get of jail free card so to speak,
and thats why they transfer "the license" to the person prompting Claude so they wash their hands of any responsibility.

So Claude is off the hook, the person submitting the code doesnt know where the code comes from.
But the Freebsd foundation would get sued for using the copyright code.

There is no way a project can determine if the AI generated code has been scraped from copyright sources or not.

Thats probably why as eseipi pointed out that the QEMU project refuses any AI generated content
because their legal team has told them of the risks.

 

[ZioMario]

Sure, you can call me a Luddite as much as you want. But again, how do you plan to bypass the FreeBSD contribution guidelines ?

I don't want to bypass anything. I want to create what I think could be nice and useful for me and for someone. First of all I want to enjoy the project by myself. Then,if someone wants to grab the code and try it on his / her machine,I will be happy. I don't have high expectations, I'm not a contributor of large projects and I'm not a programmer. But I have a good imagination and creativity. By the way, I'm finishing another project that will help people to solve another set of problems that FreeBSD suffers from (not because it's a shitty system, but because it's still anchored to the server world and it needs to grow for home use). So,stay tuned.

 

[ZioMario]

I'm sure you caught the recent news that QEMU is revising its policy to start accepting AI / LLM-generated code. What do you think about that ?

https://lists.nongnu.org/archive/html/qemu-devel/2026-05/msg07614.html

so,my work for accelerating a QEMU vm with bhyve / vmm becomes totally acceptable. Because QEMU accepts it and because the code written by Corvin is covered by the BSD license. Maybe one day even the FreeBSD developers do the same as the QEMU ones. Should progress be hindered or encouraged ?

 

[eseipi]

What do you think about that ?

I think they can do whatever they decide with their project. This is open source.

so,my work for accelerating a QEMU vm with bhyve / vmm becomes totally acceptable.

Does it? Did you read the patch behind the link?
And by the way, do you think it's fair to claim ownership of a work one cannot even comprehend?
Because, in my opinion, it clearly isn't.

Maybe one day even the FreeBSD developers do the same as the QEMU ones.

LLM contributions to FreeBSD are allowed, you basically just have to understand what you're submitting. The bar is pretty low, if you ask me.

Should progress be hindered or encouraged ?

If you'd read this book, you at least wouldn't be asking such oversimplified questions about progress.

 

[bgavin]

Ugh.
I don’t have a dog in the AI coding event.

My problem with AI is its malignant application for snooping and marketing.

I’ve left windows after Win10 and moved to MX Linux for my daily driver because of the Win11 snooping. Yes it can be avoided with a custom install but this doesn’t apply to the mass market getting raped by Windows Recall and client side surveillance.

If I can pass thru my MX dual nVidia to a VM running a lower version of Windows (7 ?) I will be a most happy camper.

 

[ZioMario]

And by the way, do you think it's fair to claim ownership of a work one cannot even comprehend ?

Oh no. I clearly said everywhere that the bhyve code has been written by Corvin and also that the starting code of vmm that accelerates QEMU has been written by Abhinav Chavali. And that the additional code has been written by Claude.

Today there is the tool that allows to bring new ideas to life and I want to use it. And it is not even free for me. The usage of Claude is not free. It costs money. Maybe I deserve a thank you.

The ideas,the testing on top of real hardware,the time that I spend (time is money) are mine. I have some generic understanding about if and how my ideas are satisfied. Otherwise, I wouldn't have been able to complete the many AI projects I'm working on. I've only released a small portion of them on this forum. But I'm designing an entire FreeBSD-based phone by myself with Claude.

I've been playing with PCs and operating systems since the 1990s. I've accumulated knowledge and a method, which, even if they're not very well structured, they make me happy.

My way of seeing things is more pragmatic. I wanna create new tools and more opportunities for me and for everyone. That's what it counts more. At least for me. Developers have and will have fresh and interesting almost working code more and more in the future. They should be happy. No one should wait years,decades to create what has always been historically missing in the *BSDs. They should only find the time to look inside it and to make it better.

 

[ZioMario]

Netbsd-Users-List

Seems to me, QEMU is taking a measured approach to AI contributions.

LLM's code should not be reflexively banned, IMHO. This 'stuff that could be easily reverted' QEMU approach seems reasonable to me.

https://lists.nongnu.org/archive/html/qemu-devel/2026-05/msg07614.html

This is an interesting point. The QEMU developers decided that the responsibility belongs to the code providers. No one can say it's Claude's fault. I would never say that. I know Claude is a tool. But I imagine there are those who do.

 

[eseipi]

Oh no. I clearly said everywhere that the bhyve code has been written by Corvin and also that the starting code of vmm that accelerates QEMU has been written by Abhinav Chavali. And that the additional code has been written by Claude.

Absolutely, but I was responding to a specific statement.

And it is not even free for me. The usage of Claude is not free. It costs money. Maybe I deserve a thank you.

You certainly do. From Anthropic stakeholders

Developers have and will have fresh and interesting almost working code more and more in the future. They should be happy. No one should wait years,decades to create what has always been historically missing in the *BSDs. They should only find the time to look inside it and to make it better.

Yeah, right. Artists of all kinds, too, should be grateful and happy that now there's overwhelmingly more almost-beautiful art around us and that they no longer have to put as much time and effort into creating it as they used to. All that's left is the most fun part - to be a peripheral of a machine. And how efficient that is for consumers! The same goes for doctors, lawyers, teachers and many other occupations, including software developers. Wait, did I mention the economic, ecological and social benefits? They're endless! Everybody should love AI, that's for sure.

And, of course, all of that is inevitable. This is just how progress works. Embrace it or step away, don't be a loser. In the end, we'll be glad it happened. We always do.

It was my last off-topic post here, I've taken this too far already. My apologies for derailing the discussion.

 

[ZioMario]

Yeah, right. Artists of all kinds, too, should be grateful and happy that now there's overwhelmingly more almost-beautiful art around us and that they no longer have to put as much time and effort into creating it as they used to

That's wrong. Humans will become obsolete when they decide to stop producing their knowledge, NOT when new tools are invented. Strong human beings don't succumb to new technological challenges; they ride them, they use them to carve out new operating spaces, improving and reinventing themselves, with or without them.

And how efficient that is for consumers !

Efficiency isn't the only factor to consider. There are many others, and they're all ours, they don't belong to AI's. But from what you say, I understand the real problem is convincing ourselves that the only way to survive is to obey to the laws of the market.

Another problem is also having hardened one's habits so much that one is no longer able to exercise a divergent thought from that which is socially trumpeted everywhere.

 

[bgavin]

Give the politics a rest.
Nobody here wants to hear that bullshit.

 

[ZioMario]

What I'm trying to say is that with every great human invention, there have always been those who took it badly, as if it were a sign of the end of the world,(or of their world ?). But then they survived. No one dies 'cause this. People don't die 'cause technological progress. They die from wars and epidemics. Technology is good as long as it isn't used for military purposes used against us.

 

[First_Law_of_Unix]

I work in AI LLM engineering. Most AI scientists in the industry are so stupid that they themselves say they have no idea how and why AI works. That is because they actually never touched nor learned how the low level code of it actually works. They R&D by writing all this AI theory in Python crap and hope for the best that the Python to CUDA compilers do the magic.

I could literally make my own AI startup company with my own custom scratch made LLMs using high quality large custom training datasets. Anyone who understands it ends up doing it.

Real MVPs are those who specifically do CUDA/GPU programming for AI research. In America, there is no such talent for this. They all write AI in Python. China has plenty of low level AI programmers. Most of them are PhDs. That is why China has better AI science. America's AI is crap. I have been doing this for over 7 years. If you write your AI in low level, you know why, how, when and what you are doing. When you write it in Python, you jack know nothing.

To make things even worse, nobody even understands the basic theory of how exactly LLMs work. The experts who explain it do not even know what they are talking about.

Anyone doing bare metal LLM engineering will tell you that all of it is really stupid. It becomes smart when you keep making it appear to be smart (what everyone in the industry is doing), when really it is stupid.

Thus, never trust any expert in the AI industry saying AI is smart. Believe me, it is really stupid. I have not seen any papers yet describing a more human like thought process. It is not easy to come up with one. The reason why AI is kinda smart today is due to the paper called "Attention Is All You Need" made by a group of Google PhDs trying to figure out how to make stupid AI smarter. Anyone saying AI is smart is saying so to get investment funding. People are so clueless about how LLMs work that they really do not care and just believe it like candy.

With that said, never trust what AI writes. It is all good for theory but terrible for production or something that you can trust. Use it as a tool, not as a real educated or experienced programmer. AI is improving greatly every 6 months, but there is still nothing ground breaking yet. All that Claude hype is BS. All you need is massive compute, funding, and to understand how and why. Now anyone can have an LLM that is better than Claude's military/gov-only AI crap. Never forget LLMs works because it was spoon fed to memorize it's training data. This is why LLM are stupid-expensive because it doesn't actually learn from it, it basically sees patterns and memorizes it. Thats why it needs tons of training material of the same crap just to understand it. When it requires to do something completely new which it was never trained on, it'll be hella stupid and requires making your own LLMs if its too complex or expensive for current LLMs to comprehend.

BTW I am on this thread mainly because it seems `ivshmem` got working on Bhyve. If so, congrats on that. I would like to know more. (Yes, I asked LLMs to see if `ivshmem` got working on Bhyve, its too stupid to know how to get it to work (obviously) but pointed me here lol).

 

[ZioMario]

When it requires to do something completely new which it was never trained on, it'll be hella stupid and requires making your own LLMs if its too complex or expensive for current LLMs to comprehend.

I've asked Claude several times to do something completely new and it did and what it did it works...at least for me.

-> qemu + bhyve + nvidia GPU passthru on X64 : it didn't exist before me. It works.

-> qemu + bhyve on the radxa zero 3W : work in progress,but it did most of the code correctly (Fable 5 validated the code written until now by Opus 4.6). I think the correct final working is close.

-> the porting of the Linux driver for the internal wi-fi chipset of the radxa zero 3W : it didn't exist before me. It works like a charme. it created a native driver from scratch.

BTW I am on this thread mainly because it seems `ivshmem` got working on Bhyve. If so, congrats on that. I would like to know more. (Yes, I asked LLMs to see if `ivshmem` got working on Bhyve, its too stupid to know how to get it to work (obviously) but pointed me here lol)

I have a good relationship with Claude. I think it saved in its memory the projects that we have been able to achieve together and when someone asks if a specific feature has been added or not,it remembers that I belong to this forum and it suggests to come here.

 

[First_Law_of_Unix]

I have used all the top frontier AIs since 2022, and not a single one can comprehend a single shot that is optimized for high performance. On a lucky day, it will give working code with one shot, but it will never be optimized. It requires lots of babysitting and direction to get the best optimization from the original code it initially gave, which is something firmware engineers are specifically paid to do to begin with. There are also many slick ways to get something optimized (only known by industry experts), but many times AI will never mention it because it memorized the methods everyone commonly uses.

For instance, I asked AI to provide low latency ways to dump webcam frames from a Linux VM into a FreeBSD host (using byhve). It gave me all the crappy methods, and I had to ask the same question again (by the way, this is actually teaching it just by prompts) to get more advanced methods by mentioning DMA. This alone proves that AI is not thinking intelligently, and I am rather teaching it how to think for me.

With that said, I always hear good things about Claude. I personally did not like it when I first used it last year. It failed to comprehend helping with a custom high performance low latency UVC USB driver using the Windows WinUSB framework. I only got it to work how I wanted by asking old school Windows driver programmers at a forum.

I have a good relationship with Claude. I think it saved in its memory the projects that we have been able to achieve together and when someone asks if a specific feature has been added or not,it remembers that I belong to this forum and it suggests to come here.

There are many papers introduced at the beginning of this year about various ways AI should store all user conversations to better understand the user and their work. Gemini AI has already done this, and it sucks. However, Grok AI mentioned your work. Frontier AI models are now implementing real time web searches using the RAG approach to find information that closely matches user requests. They are also constantly scraping the entire internet, making massive training datasets from it, and training AI on it.

I am interested in how you got ivshmem working with bhyve. It will save me maybe a good week or two. What work did you do that I should start looking into? Knowing you did all of this with AI, I can assure you with high certainty that it will have many flaws (I hope it's not big) in optimizations that I need to fix so that its actually meaningful and useable for my low latency applications.

For me (as well as top tech companies), any hardware or low level code must be perfect in terms of the highest efficiency and performance. If the code is not optimal, that is billions of dollars of money lost in energy and a waste of compute. This is why frimware/hardware engineers gets paid much more than software engineers and AI engineers. I use AI to help perfect things in this space, but it can never ever get it right on the first try in terms of optimization (not yet lol).

 

[ZioMario]

For instance, I asked AI to provide low latency ways to dump webcam frames from a Linux VM into a FreeBSD host (using byhve). It gave me all the crappy methods, and I had to ask the same question again (by the way, this is actually teaching it just by prompts) to get more advanced methods by mentioning DMA.

thanks for remind me this. The behavior you mentioned sounds strange for me because when I asked Claude to enable the audio from the Waveshare display that I've bought for the FreeBSD based phone that I'm building,it offered me the DMA as a better method for producing a clearer and stable audio path. I simply accepted and now I have it. Maybe because I've got the habit to ask it to read how has been written the Linux code before to start writing the implementation for FreeBSD,so it is inspired by the code that you offer to it as a starting working base and it takes the working code base as a model to be inspired to make a new implementation. Acts this way and let me know.

I am interested in how you got ivshmem working with bhyve. It will save me maybe a good week or two. What work did you do that I should start looking into

I asked Claude to explain me how could I have replicated the functioning of VFIO on FreeBSD. That is, to write an implementation as close as possible to VFIO for Linux. Simple. It offered me ivshmem.

I have used all the top frontier AIs since 2022, and not a single one can comprehend a single shot that is optimized for high performance. On a lucky day, it will give working code with one shot, but it will never be optimized.

Do you want ready-made "food" ? The fact that the code produced by Claude is not optimized is a valid reason for a programmer to optimize it with its help. I think humans should cooperate with AI,not that they let them do all the work OR that they don't want to be helped at all by AI. Every extreme position is not good.